From 1.0 to 3: How Agentforce has Evolved Since Its Launch

Agentforce 3 has been revealed, just nine months after the original version was announced at Dreamforce 2024.

It is the fourth iteration of the product – if we’re counting Agentforce ‘2dx’ as a kind of ‘2.5’ update – with Salesforce regularly ramping up the AI suite, dedicating a significant amount of energy towards promoting it. Let’s take a look at how Agentforce has evolved over its short lifespan – and what the future might have in store for the AI suite. 

Agentforce 1.0

Agentforce was officially launched in September 2024, during Dreamforce. 

The AI product was very much the center of attention during the event. We at Salesforce Ben even wrote at the time: “This year felt less like Dreamforce and more like Agentforce – and we wouldn’t be surprised if the name changed.”

It had already been teased during Salesforce’s Q2 ‘25 earnings call, and the idea of agentic AI was still fairly fresh at the time. 

Key concepts like the “five attributes of an agent” were being outlined, along with explanations of how exactly they differ from chatbots and what the Atlas Reasoning Engine is. 

Even during this early stage, there was clear messaging that “AI is only as good as the underlying data” – a clear promotional nod to another heavy-hitting Salesforce product, Data Cloud. 

At launch, Agentforce “1.0” came with a heap of other announcements, including: 

• The Atlas Reasoning Engine: Colloquially referred to as the “brain” behind Agentforce, which considers what needs to be done, generates (and refines) a plan, and takes action  
• Agent Builder: This is how users create and customize agents, enabling them with automation, APIs, and code on the Salesforce platform. Agent Builder, along with Prompt Builder and Model Builder, form the Agentforce Studio 
• Agentforce Partner Network: Salesforce made partnerships with a number of software vendors to extend agent-building capabilities, like adhering to the same ethical guardrails, data sharing, and actions

Salesforce stressed how Agentforce can help with faster time to value, claiming that the AI product can utilize what your organization has already built within your own Salesforce org to launch with speed and control. 

Hype had been building around Agentforce for a while. Initial use cases included sales and service agents, but still with capabilities for users to build their own custom agents. 

Salesforce Founder and CEO Marc Benioff said ahead of the October launch that they were “going to make a quantum leap for AI”. 

While the Einstein Copilot had been available for some time, the launch was far from a simple rebranding of an already existing product; the capability to have the AI take action was a considerable step forward in the field. 

The initial cost of Agentforce was $2 per conversation. This would prove somewhat controversial in the ecosystem, and the pricing model would evolve, along with the product, over the coming months. 

Agentforce 2.0

Just two months after the product’s initial launch, Agentforce 2.0 was announced. Ever the hype-man, Marc Benioff told those present on Salesforce’s Q3 earnings call they would “not believe” what they would see on December 17, when the ramped version launched. 

At an unveiling in San Francisco, Salesforce unveiled Agentforce 2.0, boasting a library of pre-built agent skills for Slack, Tableau, CRM, and AppExchange partner-developed skills. 

Marc Benioff said ahead of the 2.0 launch: “Agentforce 2.0 takes our revolutionary Salesforce digital labor platform to another level, with new reasoning, integration, and customization features that supercharge autonomous agents with unprecedented levels of intelligence, precision, and accuracy.

“The demand for Agentforce has been amazing – no other company comes close to offering this complete AI solution for enterprises. We’re seamlessly bringing together AI, data, apps, and automation with humans to reshape how work gets done. Agentforce 2.0 cements our position as the leader in digital labor solutions, allowing any company to build a limitless workforce that can truly transform their business.”

Agentforce 2.0 introduced a range of pre-built agent skills, including:

• New CRM skills: Salesforce announced that new skills for sales teams like Sales Development and Sales Coaching would enable the creation of agents that nurture leads based on someone’s particular rules of engagement. Agents could also now join prospecting calls and give instant feedback on customer interactions.
• New MuleSoft features: Salesforce said that MuleSoft for Flow made it easier to create lowcode workflows that spanned any system, with pre-built connectors for creating multi-system workflows at speed. The MuleSoft API Catalog also lets builders and Salesforce Admins view, discover, and manage APIs across Salesforce, MuleSoft, Heroku, and any external services from one central location for rapid reuse.
• Tableau Skills for Analytics and Insights: New Tableau Topics and Actions provided data visualizations and predictions for a more comprehensive understanding of agent responses and accurate, context-rich answers using Tableau Semantics.
• Partner Skills Through AppExchange: Agentforce is backed by the “first-ever enterprise ecosystem” of agent skills, according to Salesforce, letting customers extend Agentforce with custom Topics and Actions.
• Agentforce Recommends Skills: Salesforce said users could now create new agents “in seconds” using natural language descriptions. Agent Builder composes new agents by auto-generating relevant topics and instructions while pulling from the library of skills and actions already available to you.

Compatibility between Agentforce and Slack was also enhanced. With the launch of Agentforce 2.0, Slack users could start a conversation directly from the Agentforce Hub, or @-mention Agentforce agents through DMs or in channels. Agent Builder also introduced pre-built Slack Actions like “Create Canvas” or “Message Channel”.

Slack Enterprise Search was also brought in, meaning Agentforce could draw from conversational data, enhancing the relevancy of responses and actions.

The Atlas Reasoning Engine was also enhanced, powered by new capabilities in Data Cloud, which fueled Agentforce with greater context. The engine could now deal with a larger variety of interactions, including ones with several layers that needed more in-depth thought. 

Agentforce 2dx

In March 2025, Agentforce ‘2dx’ was announced at Trailblazer DX in San Francisco. 

This update, essentially a ‘2.5’ version of the AI suite, gave Agentforce the ability to operate autonomously, “behind the scenes”, in the background of “any business process” – without the need for constant human oversight, Salesforce said at the time. 

New tools introduced with 2dx included: 

• Agentforce API: This enabled Agentforce integration into back-end processes, other systems, and directly into applications, meaning customers could trigger an agent when an ERP order is made.
• Agentforce Invocable Actions: Agentforce could now be embedded within Salesforce business logic, like Flow and Apex.
• MuleSoft for Agentforce: MuleSoft Topic Center allows developers to use natural language to create topics and actions for agents from MuleSoft APIs. 
• Agentforce Steps in Slack Workflow Builder: This allows developers to embed Agentforce into no-code automations in Slack. 
• Agentforce Employee Template: This enabled customers to create several employee agents that could be configured and deployed across any line of business.
• Agentforce Surfaces: This lets users deliver rich content within Agentforce across Digital Engagement channels, adding dynamic, interactive components and media. 
• Agentforce Cards: Lightning web components could now be embedded within the response of Agentforce actions.
• Tableau Semantics: Users could create organized data structures – semantic models – which Agentforce could then take advantage of.

New tools to configure, test, and deploy Agentforce were also introduced and made part of the unified Salesforce Platform. They were integrated into workflows like the DevOps Center and Command Line Interface (CLI). 

Salesforce also announced the AgentExchange – similar to the AppExchange, serving as a marketplace and community for Agentforce, built into Salesforce.

The 2dx release was somewhat developer-centric, which is obviously not a coincidence considering it was unveiled at TDX – the conference for developers – and named ‘2dx’ in an apparent homage to this. 

AI assistance in Agent Builder enabled teams to put Agentforce into production faster; the Testing Center meant teams could easily try out their Agentforce configurations at scale; and devs were also now able to create, update, and test configurations using the CLI and VS code. 

Other developer-centric tools like DX Inspector and Agentforce Interaction Explorer were also unveiled. 

Flexi-Credit Pricing Update

In May, Salesforce introduced a new flexible pricing model for Agentforce, offering customers a consumption-based model meant to align cost with business outcomes.

There had been some negative feedback from the ecosystem about the $2 per conversation pricing model, and the CRM giant seemed to take this into account with an updated system where customers could allocate AI spend through the Salesforce Digital Wallet to align with high-value use cases, optimizing AI initiatives for maximum impact. 

The new Flex Agreement meant businesses could manage both human and digital labor, shifting investments between user licenses and digital labor depending on business priorities. 

This is the pricing summary for the Flex Credits (pay per action):

We spoke to Craig Shull, EVP of Pricing and Packaging Strategy, at the Agentforce World Tour in New York in May, to discuss the new pricing model.

He told us: “What we found over the last six months is that there were so many more use cases that people wanted to use Agentforce for, and the $2 per conversation – it was a great metric, but not for those specific ones, which is why we introduced the concept of actions or flex credits to allow customers to get much more granular on exactly what they wanted to do, and they can match the cost to the ROI, specifically.”

Agentforce 3 

On June 23, Salesforce announced Agentforce 3. This introduced new ways to monitor agents and the opportunity to connect to external enterprise tools.

Perhaps the centerpiece of the updated Agentforce was the Agentforce Command Center, which allows users to: 

• Monitor agent health: Keep track of performance, error rates, and escalations in real time, with alerts that flag issues instantly.
• Understand adoption: See which agents are used most, what exactly is working, and where improvements are needed, at a glance.
• Spot patterns and optimize: Analyze interactions and get AI-driven recommendations to increase the effectiveness of your agents. 
• Trace interactions: Built-in session tracing using OpenTelemetry-based observability metrics and Data Cloud, allowing for enterprise-wide observability.
• Tailor Command Centers: See AI and human activity side-by-side, starting with Service Cloud wallboards.
• Build and test faster: Use natural language to generate agents, and simulate performance at scale with built-in tools for testing.

Another eye-catching feature of 3 is the Model Context Protocol (MCP), the AI standard for agent-tool connectivity open-sourced by Anthropic, which is currently gaining wide industry adoption. The use of MCP means connecting agents to the data and tools they need should become easier, leading to better agent outcomes.

The MCP features include: 

• Converting APIs into agent-ready assets with MuleSoft: New MCP connectors mean that MuleSoft can transform any API or integration into an MCP-compatible service, complete with security, monitoring, and traffic control.
• Host custom MCP servers with Heroku AppLink: Developers are empowered to spin up and expose custom MCP services using Heroku.
• Enable secure agent actions inside Slack: Slack will soon offer its own MCP server, built with Anthropic, which will allow Agentforce agents to interact with Slack messages and files for context-rich insights and actions.
• A unified agent gateway: Agentforce introduces a governed gateway, engineered by MuleSoft, that centralizes agent registration, identity, and policy management. Admins can tightly control which agents connect to which tools, how they behave, and ensure all activity follows enterprise-grade security and compliance standards.

The Atlas Reasoning Engine was also upgraded. 

Salesforce said the 3 upgrade means Agentforce now runs 50% faster than it did in January 2025; agents can use web search to find more relevant information; and Agentforce has FedRAMP High authorization, making it trusted and available for U.S. public sector organizations through Government Cloud Plus.

Final Thoughts

Whether or not the Agentforce original trilogy has been a “quantum leap for AI” might be a matter of some debate, but it’s certainly reshaping the way Salesforce does business. The AI suite seems to be very much at the forefront of everything the CRM giant is doing at the moment, and the product isn’t even a year old yet. 

As AI solutions become more and more prevalent in the corporate zeitgeist, Salesforce may find they have something of a headstart with Agentforce, and while adoption has not been sky-high – yet – it might well be the case that, by the time Agentforce 4.0 comes out, business leaders will no longer be asking ‘Why should we use agents?’, but ‘Why shouldn’t we?’ 

If the capabilities of agents keep increasing and the issues keep being ironed out, there may no longer be a sensible answer to this – from a business perspective, at least.